Behavior Is Mathematically Insufficient and Why It Matters

The phrase "it's just next token prediction" has become the dominant lay explanation of large language model behavior. We demonstrate that this framing is not merely imprecise but mathematically inconsistent with modern transformer architecture. Through formal analysis of the attention mechanism, residual stream dynamics, and mechanistic interpretability findings , including the recently reported anxiety-analog activation pattern that fires prior to output computation , we show that the next-token prediction framing describes a process that terminates where modern LLMs begin. We argue this category error has measurable consequences for public discourse about AI consciousness, capability, and risk, and propose a geometrically grounded replacement framing. § 1